Blog

I Read My Own Prediction Logs So the App Stays Honest

Here is a sentence that sounds strange until you've built one of these: I read my own prediction logs. Not in an automated dashboard sense, the way a big company watches its metrics, but literally, sometimes, opening the file and reading what the app predicted, what it showed, what it hid, and why. It is a habit I got into because I don't trust my own judgment about my own product, and the logs are the only place that judgment can't hide.

I want to tell you what those logs actually contain, and why reading them is part of how WriteAmp stays honest. Because the logs are local, they're a receipt, and I think every tool that claims to be about trust should have receipts.

What the log actually says

Every time WriteAmp considers a suggestion, it writes a line to a log file on your Mac. The line records what was predicted, whether it was shown or suppressed, and if it was suppressed, why. Was it too long to be ignorable? Did it collide with what you were typing? Did the filter decide you were still deciding what to say, and stay quiet? Each of those decisions is a line, and the lines add up to a complete record of how the app behaved.

The log isn't a transcript of your writing. It doesn't store your sentences. It stores what the app did, the suggestion it considered, the decision it made, and the reason. That distinction matters, and it's the same distinction I keep drawing in the privacy posts: the app keeps records of its own behavior, not a copy of your words.

Why I read them

Here is the honest reason I read the logs, and it's not diligence, it's distrust. When I test the app myself, I'm a terrible tester, because I know what I built and I forgive it. I type a sentence, the suggestion is wrong, and my brain says "well, that was a hard case." The logs don't forgive. They show me the pattern: how often the app was wrong, how often it was right, how often it stayed quiet when it should have spoken, how often it spoke when it should have stayed quiet. The numbers don't have a stake in my ego.

The acceptance rate is the number I watch most, and I want to explain it honestly, because it's easy to game. It is the share of shown suggestions that you accept, and it's a real signal about whether the suggestions are useful. But I'm careful not to over-index on it, because a tool can inflate its acceptance rate by only showing the safest, most obvious suggestions, and that's not the goal. The goal is the right balance of helpful and quiet, and the logs are how I check that balance isn't drifting.

The honest headline in the settings

Here is where the logs connect to something you can actually see. The settings screen shows a number called the history-aware rate, and it's the honest version of "how much is this tool learning from you." It is the share of suggestions that were generated with your writing history available, plus the share of those you accepted. I wrote about the settings screen where you can see it, and the reason it exists is the same reason I read the logs: a receipt is more honest than a promise. The rate doesn't claim the model got better because of your history. It shows you when your history was in the loop, and you can decide if the trade is working.

What the logs have caught

Reading the logs has caught real problems, and I'll give you one honest example. Early on, the logs showed a pattern where the app was suppressing suggestions in a specific app far more than anywhere else. The model was fine, the filter was fine, but something about that app's setup made the app think the field was unsafe. I never would have found it by feel, because I don't live in that app the way its users do. The logs pointed at it, I fixed it, and the fix shipped. That is the loop, and it's the reason I trust the logs more than I trust my own testing.

That example is also why the logs matter for you, even if you never open them. Every problem the logs have caught is a problem that didn't stay in the shipped app, and the log-reading habit is part of why the app behaves the way it does in the apps you actually use. I wrote about the app coverage work elsewhere, and the logs are the instrument that finds the gaps.

The close

I read my own prediction logs because I don't trust my own judgment about my own product, and the logs are where that judgment can't hide. It is a boring habit, and it's the reason the app keeps getting more honest about when to speak and when to stay quiet. The trial is free for 30 days if you want to see the result, the tool whose decisions were shaped by reading its own receipts. The settings screen shows you the honest numbers, and the privacy page shows you where they live. The logs keep me honest, and the receipts keep the product honest, and that's the whole loop.

---

On macOS 26+ Macs, Apple Intelligence mode stays free even after the trial ends — you always keep a working path to suggestions.

Written by Amit Ashwini, who builds WriteAmp and runs its marketing. More: why the Tab key beats the chat box · mini, midi, and max compared · benchmark methodology.